Learning about Learning in Games through Experimental Control of Strategic Interdependence

We conduct experiments in which humans repeatedly play one of two games against a computer decision maker that follows either Roth and Erev's reinforcement learning algorithm or Camerer and Ho's EWA algorithm. The human/algorithm interaction provides results that can't be obtained from the analysis of pure human interactions or model simulations. The learning algorithms are more sensitive than humans in calculating exploitable opponent play. Learning algorithms respond to these calculated opportunities systematically; however, the magnitude of these responses are too weak to improve the algorithm's payoffs. Human play against various decision maker types does not significantly vary. These results demonstrate that humans and currently proposed models of their behavior differ in that humans do not adjust payoff assessments by smooth transition functions and that when humans detect exploitable play they are more likely to choose the best response to this belief.